Table 5 Hyperparameter summaries for ten common baseline models.
Model | Core approach | Typical hyperparameters |
|---|---|---|
1. CNN-Baseline | 2–4 conv layers + FC | LR \(\approx 1\times 10^{-3}\), batch=16, dropout=0.2 |
2. RNN-Baseline | BiLSTM/GRU (2 layers) | LR \(\approx 5\times 10^{-4}\), seq length=100, focal(\(\gamma =1\)) |
3. CNN-RNN Hybrid | CNN + LSTM stack | LR \(\approx 2\times 10^{-4}\), Weighted CE |
4. CRNN-Attention | CNN + LSTM + Attention | LR \(\approx 1\times 10^{-3}\), heads=4, dropout=0.2 |
5. TCN Model | Temporal Conv Net | LR \(\approx 1\times 10^{-3}\), kernel size=3, Weighted CE |
6. RNN-Transformer | LSTM + small Transformer | LR \(\approx 5\times 10^{-4}\), 2 layers each |
7. Multi-Task CNN | Shared trunk + tasks | LR \(\approx 1\times 10^{-3}\), Weighted CE + Dice loss |
8. DenseNet-Style | Dense blocks + global pooling | LR \(\approx 5\times 10^{-4}\), batch=8, dropout=0.2 |
9. Transformer-Lite | 1–2 self-attention layers | LR \(\approx 2\times 10^{-4}\), heads=2, focal(\(\gamma =2\)) |
10. Wavelet-CNN | Wavelet-based conv filters | LR \(\approx 1\times 10^{-3}\), Weighted CE, L2 reg |
